Object Recognition for Robotics Based on Planar Reconstructed B-Rep Models

Abstract

Current robotic systems in small and medium-sized enterprises as well as in households need a perception system to enable an interaction of the robot with the environment. One necessary ability of a perception system is a method to recognize objects in a given scene. We contribute an approach to object recognition based on the input of an eye-in-hand depth camera, which uses boundary representation (B-Rep) models as the internal data format. Our method applies three steps: generate B-Reps from point clouds, generate and select hypotheses from an object database, as well as build up a world model. In the evaluation, we discuss the classification rate and conclude, that our approach is suitable for robotic applications, as it is an anytime, surface-based method without the need for a prior object segmentation.